Compute & chips

Why is everyone suddenly fighting over computer chips? Because the bigger a model gets, the more raw computing power it takes to build and run — and that has become the whole game's bottleneck.

Bigger models, hungrier machines

Training a model means running an unthinkable number of calculations over mountains of data. Compute — that raw processing power — is the fuel. Scale a model up and its appetite for compute doesn't creep up; it explodes. Frontier models take thousands of specialised chips running for weeks to train, and a serious slice of a datacentre just to answer you.

The chip gold rush

Those specialised chips are GPUs, and one company — Nvidia — makes most of the best ones, which is why it became one of the world's most valuable firms. Google, Amazon and others now design their own too. Access to chips increasingly decides who can build frontier AI at all: it's less about clever code, more about who can get the hardware and the power to run it.

The short version

• —Compute — raw processing power — is what trains and runs a model.
• —Scaling up multiplies the chips and energy needed, fast.
• —GPUs (mostly Nvidia's) are the scarce resource everyone's chasing.